hunter paulsonprojects |
|||||
| blog | art | projects | contact | ||
can LLMs actually navigate a website that is in their training data?
a benchmark for agentic web navigation where LLMs compete to navigate from one Wikipedia page to another in the fewest number of steps.
while building wikiarena.org I got nerd sniped into building what may be the fastest wikipedia race solver on the internet.
Can you type faster than GPT-2?
runs GPT-2 locally in-browser using WebGPT and generates typing tests using only words in the selected wordset.
I built this to learn local llm inference and also so I could practice typing probable word sequences instead of words sampled uniformly at random.
creator and head of PayPal’s internal cloud agent harness ‘Pal’ and underlying infrastructure.
designed the algorithm and built the infrastructure to identify the root-cause component of every HTTP 4xx and 5xx in real time. data was used to identify and triage issues in the payment flow, reclaiming $40M in revenue. equivalent to ~1% growth of branded checkout in 2025.
helped build real-time metrics pipelines for every http response and server log across all PayPal brands. the single, structured, source of truth for all site reliability metrics.
wrote transpiler from python to bigquery sql.
automated data lineage documentation by parsing and graphing every OLAP sql query ran in the company daily.
trained computer vision models and built data labeling pipelines for assemply line automation
contributed to the open source flight software framework that supported the Mars Helicopter Ingenuity.
fun fact: I ran the autoformatter.
worked on the Phoenix 3U Cube Satelite. launched from the ISS on Feb 19, 2020.